Fraud Detection in E-commerce Using

Natural Language Processing

by

Iftekhar Kabir
18201106
Marium Khan Momo
18301069
Tahsin Tazrian
19101520

A thesis submitted to the Department of Computer Science and Engineering

in partial fulfillment of the requirements for the degree of
B.Sc. in Computer Science

Department of Computer Science and Engineering

Brac University
January 2023

© 2023. Brac University

All rights reserved.
Declaration
It is hereby declared that

1. The thesis submitted is our own original work while completing degree at Brac
University.

2. The thesis does not contain material previously published or written by a

third party, except where this is appropriately cited through full and accurate
referencing.

3. The thesis does not contain material which has been accepted, or submitted,
for any other degree or diploma at a university or other institution.

4. We have acknowledged all main sources of help.

Student’s Full Name & Signature:

Iftekhar Kabir Marium Khan Momo

18201106 18301069

Tahsin Tazrian
19101520

i
Approval
The thesis/project titled “Fraud Detection in E-commerce Using Natural Language
Processing” submitted by
1. Iftekhar Kabir (18201106)
2. Marium Khan Momo(18301069)
3. Tahsin Tazrian (19101520)
Of Fall, 2022 has been accepted as satisfactory in partial fulfillment of the require-
ment for the degree of B.Sc. in Computer Science on January 17, 2023.

Examining Committee:

Supervisor:
(Member)

Dr. Farig Yousuf Sadeque

Assistant Professor
Department of Computer Science and Engineering
Brac University

Program Coordinator:
(Member)

Dr. Md. Golam Robiul Alam

Professor
Department of Computer Science and Engineering
Brac University

Head of Department:
(Chair)

Sadia Hamid Kazi

Chairperson
Department of Computer Science and Engineering
Brac University

ii
Abstract
Electronic commerce sometimes referred to as e-commerce is a type of business
that enables both businesses and private individuals to purchase and sell products
and services online. E-commerce in Bangladesh is thriving from the last decade,
especially during the coronavirus pandemic with the growth of online sales. Digital
commerce is currently struggling to regain trust after allegations of annexation and
fraud surfaced against a few firms in recent months. Over 11.48% clients of the
internet business area were beguiled last year from di↵erent web based business
and Facebook trade (business) sites. Fake reviews are one of the most prominent
fraudulent activities in this field. When we try to buy anything online or book
any hotel from an app or a ride from any ride sharing app we heavily rely on the
reviews of past customers.It makes the decision making process easier. This is why,
with the ongoing development of e-commerce platforms online reviews are seen as
essential to upholding a company’s reputation. Generally a positive feedback from
a customer gathers the attraction of many searching for the same product. For this
reason, many e-commerce sites are generating fake reviews to attract more customers
towards them. Detecting fake reviews is an ongoing research area. As all the reviews
are not trustworthy and honest, it is crucial for us to develop techniques for detecting
fake reviews. We are proposing a machine learning approach to generate and detect
fake reviews.We used Natural Language Processing(NLP) to extract meaningful
features from a text for detecting fraud reviews. Therefore, in this study, we present
a comprehensive and e↵ective framework that enhances the efficacy of fake review
identification using Support Vector Machine(SVM) and Logistic Regression machine
learning algorithm among several machine learning algorithms for detecting fake
reviews.

Keywords: Fake Review; Machine Learning; Support Vector Machine; Logistic

Regression; Detection; BiDirectional Long Short Term Memory

iii
Acknowledgement
BRAC University provided support to facilitate the completion of this study. First
and foremost, we thank the Great Allah for sparing us damage during the COVID-
19 pandemic and enabling us to finish our research on schedule. We would want
to take this opportunity to thank our supervisor, Dr. Farig Yousuf Sadeque sir for
everything that he has done to help us and for letting us work for him. And last,
many thanks to our parents for their thoughtful prayers and help.

iv
Table of Contents

Declaration i

Approval ii

Abstract iii

Acknowledgment iv

Table of Contents v

List of Figures vii

List of Tables viii

Nomenclature ix

1 Introduction 1
1.1 Research Problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.2 Related Works . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3

2 Literature Review 6
2.1 Natural Language Processing(NLP) . . . . . . . . . . . . . . . . . . . 6
2.2 Synthetic Review Detection . . . . . . . . . . . . . . . . . . . . . . . 6
2.3 Supervised Learning Method for Detecting Fake Review . . . . . . . . 7
2.4 Unsupervised Learning Method for Detecting Fake Review . . . . . . 7
2.5 Text Reuse Detection . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
2.6 Feature Engineering for fake review detection . . . . . . . . . . . . . 8

3 Background Studies 9
3.1 Computer generated reviews . . . . . . . . . . . . . . . . . . . . . . . 9
3.2 Human generated via review farms . . . . . . . . . . . . . . . . . . . 9
3.3 Human generated fake negative review . . . . . . . . . . . . . . . . . 10
3.4 Human generated fake positive review . . . . . . . . . . . . . . . . . . 10
3.5 Fake Reviews Detector: Broad Overview . . . . . . . . . . . . . . . . 11
3.5.1 Natural Language Processing techniques . . . . . . . . . . . . 11
3.5.2 Machine Learning in NLP . . . . . . . . . . . . . . . . . . . . 11
3.5.3 Support Vector Machine (SVM) Machine Learning Model . . . 12
3.5.4 Logistic Regression Machine Learning Model . . . . . . . . . . 12
3.5.5 Term Frequency Inverse Document Machine Learning Model . 12
3.6 Analysis of fake reviews using review text . . . . . . . . . . . . . . . . 13

v
 3.7 Analysis of fake reviews using reviewer behavior . . . . . . . . . . . . 14

4 Methodology 17
4.1 Workflow . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
4.2 Dataset . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
4.3 Preprocessing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
4.4 Model Implementation . . . . . . . . . . . . . . . . . . . . . . . . . . 19
4.5 TF-IDF . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
4.6 GloVe Vectorizer . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
4.7 Support Vector Machine Algorithm(SVM) . . . . . . . . . . . . . . . 20
4.8 Logistic Regression(LR) . . . . . . . . . . . . . . . . . . . . . . . . . 21
4.9 BiDirectional LSTM . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
4.10 Confusion matrix, Accuracy, F-1 Score , Recall and Prescision . . . . 22

5 Experiments and Results 24

5.1 Result analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
5.1.1 SVM and LR Results using TFIDF: . . . . . . . . . . . . . . . 24
5.1.2 SVM and LR Results using GloVe: . . . . . . . . . . . . . . . 26
5.1.3 Bidirectional LSTM accuracy and Loss function: . . . . . . . . 26
5.1.4 Discussion: . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27

6 Concluding Remarks and Future Work 29

6.1 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
6.2 Limitations and Future Work . . . . . . . . . . . . . . . . . . . . . . 30

Bibliography 33

vi
List of Figures

4.1 Workflow . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
4.2 Bidrectional LSTM . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
4.3 Confusion matrix . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23

5.1 Confusion matrix for SVM usign GloVe . . . . . . . . . . . . . . . . . 26

5.2 Confusion matrix for LR usign GloVe . . . . . . . . . . . . . . . . . . 26
5.3 Confusion matrix for SVM usign GloVe . . . . . . . . . . . . . . . . . 27
5.4 Confusion matrix for LR usign GloVe . . . . . . . . . . . . . . . . . . 27
5.5 Bidirectional LSTM Accuracy and Loss Function . . . . . . . . . . . 27
5.6 Models Accuracy Comparison . . . . . . . . . . . . . . . . . . . . . . 28

vii
List of Tables

5.1 Results of using di↵erent parameters of SVM with TFIDF . . . . . . 25

5.2 Results of using di↵erent parameters of LR with TFIDF . . . . . . . 25
5.3 Results of using di↵erent parameters of SVM with GloVe . . . . . . . 26
5.4 Results of using di↵erent parameters of LR with GloVe . . . . . . . . 27
5.5 Models accuracy comparison table . . . . . . . . . . . . . . . . . . . . 28

viii
Nomenclature

The next list describes several symbols & abbreviation that will be later used within
the body of the document

CG Computer Generated

CN N Convolutional Neural Network

LR Logistic Regression

LST M Long Short-Term Memory

NB Naive Bayes

N LP Natural Language Processing

OR Original Review

RN N Recurrent Neural Network

SSL Semi-supervised learning

SV M Support Vector Machine

TF IDF Term Frequency-Inverse Document Frequency

ix
Chapter 1

Introduction

E-commerce is becoming the most popular platform for people all over the world
because of its easy to access and user-friendly interface. E-commerce sites are being
dominated by the “phenomenon of fake”. Reviews may have an impact on customers
in a variety of industries, but they are especially important in the world of online
commerce where customers use comments and reviews to determine whether or not
to buy products and services. Generally service providers consult with their clients
to provide their valuable feedback on the items they have purchased from them so
that it can improve their reputation and make their business trustworthy to other
customers. Platforms for social networking and online shopping have assimilated
into today’s society. As a result, the amount of data on the Web is rapidly increasing,
but further research is needed to determine its quality. People now use social media
platforms to express themselves and remain in touch by talking about the news,
weighing in on political issues, viewing movies or buying things[21]. Unfortunately,
this online data can spread rapidly and easily and it is not difficult to tamper with
it. Today particularly during emergency situations like the COVID-19 epidemic that
we currently experienced, fake news and misleading information have grown more
and more obvious. Generally, in Fraud detection it wants to get to a place where
it build like a cyclical or feedback loop of sorts that gets a better fraud detection
as time passes. So what really want focus on, in fraud detection, especially when
it has to do with content or things that are readable by an NLP, which it want to
start with the idea of understanding , so that ends up being a text mining or NLP
largely natural language understanding. So, we want to have an understanding of
what the specific piece of content is saying and start understanding how it might
relate to fraud. Numerous research have shown that in order to make the COVID-
19’s curvature flatter, the curve of rumors and false information about the virus
needs to be flattened. A significant example of where phony reviews and profiles are
prevalent is on e-commerce platforms, where there is an increasing number of new
products and fierce competition amongst businesses. Online reviews can be fake, yet
they are nonetheless important for both customers and businesses. Consumers must
inspect them in order to judge the products’ quality and whether or not to purchase
them. On the other hand, businesses require internet reviews to enhance their
o↵erings and increase sales. Fact-checking systems that can recognize bogus reviews
or fake users must be created in order to improve the transparency and quality of
the data otherwise, the Online world would be a sea of unusually pervasive rumors
and false information. Clearly, this task is not a simple one. Despite significant

1
work by scholars in this field, there is still room for improvement. We must first
acknowledge the properties of such data and the fraudulent users’ behavior patterns
in order to fully comprehend the issue. Positive internet evaluations, according
to 90% of consumers who remembered reading them, had an influence on their
purchasing decisions. In addition, 86% of them claimed that unfavorable online
reviews influenced their purchasing choices[33]. Therefore, it is imperative to have a
reasoning system that can identify these reviews especially when taking into account
the following factors: 1.Customers who read fake reviews are led to believe that
they are completely informed about the goods they are purchasing[33]. 2. To
prevent customers from giving a product a positive review, some phony reviews
are written[33]. 3. Many false reviews are published to either enhance or degrade
the product[33]. 4. Because of the paucity of training data that takes into account
the reviewers’ behavioural patterns, pure supervised learning algorithms may not
be the best option. The suggested method makes use of supervised learning to
analyze the review’s sentiment, which is taken into account by one KB rule[33]
using NLP which is natural language processing to try to predict fake reviews.
Now, this is obviously something that has come to the fore more recently than
in the past. But, we venture to speculate that fake reviews has been something
in and around the globe pretty much forever there’s that famous quote that a lie
can make it around the world. Before the truth has time to put its pants on so
that has always been the case and it will probably always be the the case so the
real question is this can we utilize new technology data science modeling di↵erent
things like that to try to predict if the review is genuine or not whether it is a
hoax, etc. before or once social media starts to spread it. The detection of fake
reviews in natural language processing is a crucial but difficult problem (NLP)[32].
In addition to greatly increasing information accessibility, the quick development
of social networking sites has also sped up the dissemination of false evaluations.
Because of this, the impact of bogus reviews has been expanding, sometimes even
a↵ecting the o✏ine world and endangering public safety. Given the vast amount
of online material, automatic fake review identification is a realistic NLP challenge
that will benefit all online content producers by minimizing the amount of time and
e↵ort needed for people to recognize and stop the propagation of phony reviews.
In this study, we discuss the difficulties in detecting false reviews as well as related
issues[17]. We comprehensively examine, contrast and evaluate the potentials and
restrictions of the task formulations, datasets and NLP alternatives that have been
created for this task. We describe prospective research directions based on our
findings, including more accurate, thorough, fair and useful detection methods. We
also emphasize the distinction between the detection of false reviews and other
similar tasks, as well as the significance of NLP tools for fake review detection[17].

1.1 Research Problem

Computer science’s natural language processing (NLP) discipline examines how peo-
ple and machines communicate. Alan Turing devised an intelligence test in a 1950s
article that is now known as the Turing test. Results have been achieved in the
areas of language modeling, parsing and natural language challenges using more
contemporary methodologies, such as deep learning[24]. In order to detect false
product reviews at one of the top virtual merchants in the world, we tested state-of-

2
the-art algorithms for natural language processing as part of our research project.
All of our models were developed using the amazon reviews dataset, which provides
explicit labels identifying false reviews[29]. Deceptive opinions, spam reviews and
spam opinions are frequently used as definitions for fake reviews. Their authors
may also be referred to as spammers. The three categories of spam opinions often
known as phony reviews are as follows: 1. Users that publish reviews with untrue
comments do so in an e↵ort to harm the reputation of a company or product or
to support it. These evaluations are referred to as phony or deceptive reviews and
since legitimate and fraudulent reviews are similar to one another, it can be difficult
to identify them by reading alone. 2. Only those who comment on the brand of the
products are described in reviews of that brand. 3. Non-reviews that are pointless
don’t provide a real viewpoint or are merely ads. The last two varieties referred to
as disruptive spam opinions pose little risk and are easily distinguishable by anyone
who reads them. We must take into account the following two samples of reviews
from a Yelp Chi real-life public dataset in order to describe and comprehend the
nature of phony reviews. While the second review is bogus, the first is real. Review
1: ”I like staying here. The sta↵ is really kind and makes you feel right at home.
Fantastic location and excellent lodging.” Review 2: ”What an incredible hotel.
The employees are excellent and thoughtful. Benefits like the free bike rentals are
fantastic. The building’s refurbishment and history are both quite fascinating. I
appreciate you making my stay so wonderful with your assistance[34].”

1.2 Related Works

User reviews have a significant influence on how much money a company makes
in e-commerce. Before choosing any goods or service, online users rely on reviews.
As a result, the legitimacy of online evaluations is essential for businesses and has
a direct impact on their profitability and reputation[19]. Because of this, some
companies pay spammers to publish phony reviews. These fraudulent reviews take
advantage of consumer purchasing choices[19]. As a result, during the past twelve
years a lot of research has been done on how to spot false reviews. However, a survey
that can evaluate and summarize the current methods is still lacking. This survey
study summarizes the existing datasets and their gathering techniques in order to
address the issue and describes the task of detecting fraud review. It examines the
feature extraction methods currently in use. Additionally, it provides a compre-
hensive summary of the available methodologies and assesses them in two groups
traditional statistical machine learning techniques and deep learning approaches to
find any shortcomings. Additionally, we carry out a benchmark research to assess
the e↵ectiveness of various transformers and neural network models that have not
previously been used to the identification of fraudulent reviews. According to re-
search findings on two benchmark datasets, RoBERTa surpasses state-of-the-art al-
gorithms in a mixed domain with the maximum accuracy of 91.2% for the deception
dataset and this performance may be used as a benchmark for more study[31]. We
conclude by highlighting the research area’s present deficiencies and potential future
directions. In regards to product reviews, Jindal and Liu initially raised the issue of
bogus reviews identification in 2008. Multiple ways have been used to address the
issue, including classifying reviews as fake or genuine by examining the review con-
tent, user behavioral traits or sentiment analysis techniques. Many researchers have

3
worked to develop cutting-edge methods for determining whether reviews are real or
spam. As stated in reference, several detection levels should be taken into account;
the primary ones are as follows: review content and user behavior-based detection,
review variances among rating-based detection and review content-based detection.
Supervised models use the first two categories, however it may be challenging to
achieve correct results because training the models may need a huge dataset[35].
To improve the overall performance, the authors suggested taking into account the
third category. In-depth research has recently been conducted on the development
of machine learning-based spam detection systems. Reference presents two distinct
methods for identifying false reviews, where new semantic characteristics were ex-
tracted and classified using supervised learning methods using a real Yelp review
dataset. In contrast to n-grams, the research suggests a collection of behavioral
traits and demonstrates how include them improves accuracy. The length of the re-
views, the intervals between reviews, and the ratio of each good to negative review
were all behavioral factors that the authors said helped them achieve high accuracy.
For classification, Support Vector Machine (SVM) has been employed. A feature
framework for fake review identification has been developed using a classification al-
gorithm. The study resolves the issue by pulling data from the actual Yelp dataset in
order to produce a brand-new dataset for the consumer electronics area. The frame-
work does employ two categories of features for feature extraction: review-centric
features which are concerned with the review as a text and user-centric features
which are focused on the reviewers’ behavioral patterns, such as their personal, so-
cial and review-related activities. As well as verbal characteristics like the length or
content of the review, nonverbal behavioral traits like the number of likes or dislikes,
the average posting frequency and review updates among others can be useful in
identifying fraudulent reviews. The study proposed non-machine learning methods
that look at relationships between reviewers, reviews and content similarity, such as
graph-based or pattern matching methods. Additionally, the research in Reference
shows that incorporating the reviewers’ behavioral attributes increased accuracy by
about 20% when compared to n-grams, but they also noted that additional informa-
tion such as IP addresses, user logs or session lengths might be employed to enhance
the outcomes. Other strategies suggested a phony review recognition system by
capturing questionable review time intervals. Using private information such as IP
addresses and MAC addresses can improve the performance of the false detection
system. The issue that there are no dataset quality requirements that can be used
to attain 100% accuracy in determining whether a review is false or not is shared by
both articles. Additionally, it exhibits a variety of qualities that make it easier to
identify phony reviews. These elements, which include the reviewers, their reviews
and the capability of repeating reviews are mostly used to identify similarities in
review content. Additionally, the things that have already been examined as well as
how frequently they are done are taken into account. The direct e↵ect of internet
evaluations on the box office earnings of a particular film, indicating that consumers
occasionally use these online services to determine whether or not to watch a film.
One of the main problems when using machine learning techniques is the imbalance
between the states of the two classes, false and real information. Getting labels
for misleading information might be challenging. These are frequently gathered
manually by professionals, educated volunteers or employees of Amazon Mechanical
Turk[14]. The procedure entails a lot of manual work and the assessors might not

4
be able to categorize every piece of false information they encounter. Rule-based
procedures, in contrast are acknowledged as a white box providing traceability and
transparency for important judgments that call for a deeper level of explanation
than is typically provided by machine learning approaches. Additionally, the choice
between a rule-based system and a system using machine learning rely on the nature
of the issue at hand and mostly involves balancing the trade-o↵s between efficiency,
training costs and comprehension.

5
Chapter 2

Literature Review

2.1 Natural Language Processing(NLP)

The term ”Natural Language Processing,” which stands for artificial intelligence,
refers to a computer system’s capacity to comprehend spoken language. Computer
science’s field of ”Natural Language Processing” is concerned with human-computer
interaction. Users of NLP may ask inquiries about any topic and receive a quick
answer in a matter of seconds. This system responds to the query with natural
language responses. The system provides precise responses to the queries without
providing extra or undesired information. This method scales other language-related
activities as well as enables computers to converse with people in their native tongue.
Speech recognition, language comprehension and language production are typically
difficult aspects of natural language processing. NLG deals with generating spo-
ken or written language from unstructured data. This is perhaps the best known
database to be found in the pattern recognition literature. NLU deals with Un-
derstanding the input given by the user as a part of natural language. Basically,
it is exploring a dataset of how many rows are there in the database, how many
labels and how many data are adding and missing. Applying automatic detection
NLP approaches emphasize lexical features (i.e., textual properties) such keywords
or -s, punctuation, n-grams, latent themes and semantic similarity when processing
reviews as textual data[1] . Along with NLP we can also consider human detectors
to assist us in detecting fake reviews . Harris[2] suggested a hybrid method in which
the output of two machine learning classifiers was combined with psycho-linguistic
features that had been retrieved algorithmically.This hybrid technique improved ma-
chine performance by 0.2 percentage points, demonstrating that human input may
lead to a modest (but statistically significant) advantage over a wholly automated
strategy. Humans might concur or disagree with the computer decision.

2.2 Synthetic Review Detection

Due to the scarcity of review spam samples and the difficulties in categorizing them,
the majority of the datasets used in the earlier research are synthetically constructed
[3]. Synthetic reviews that use sentence transplants have a subtler semantic inco-
herence between sentences than natural writings. Because these synthetic datasets
are not always reflective of real-world review spam, developing and testing classi-
fiers based on them can be challenging. For instance, the derived characteristics

6
and outcomes while evaluating the artificial AMT dataset used in [4,5,6] and Yelp’s
filtered reviews dataset substantially varied, particularly in the case of n-gram text
features[7]. When these datasets’ classification performance is compared, it is clear
that using the Yelp review dataset as a benchmark, the classifier got an accuracy
of 65%, but while using synthetic reviews it achieved 87% accuracy. This 22% de-
cline in accuracy suggests that synthetic reviews vary from real-life fake reviews in
terms of di↵erentiating characteristics and that AMT does not correctly represent
real-world spam reviews in its reviews.

2.3 Supervised Learning Method for Detecting Fake

Review
A combination of machine learning and artificial intelligence supervised learning
is sometimes referred to as supervised machine learning. It shines out because it
trains algorithms that properly classify data or forecast occurrences using labeled
datasets.During the cross validation process, weights are modified as input data are
put into the model until the model is well fitted.With the use of supervised learning,
businesses may find sustainable solutions to a wide range of real-world issues. In[21]
Their proposed model is divided into four stages. The first step which involves
data collecting and data preparation uses both labeled and unlabeled datasets and
preprocesses them.Preprocessing is done using both labeled and unlabeled data in
Natural Language Processing (NLP) methods such as stop word and punctuation
removal, changing to lowercase letters in english and stemming. In the second
stage the Active Learning Algorithm is used to gradually label all of the unlabeled
data while the learner evaluates the dataset’s accuracy by comparing the probability
di↵erence with a threshold value for accurate classification. The third step covers
the technique for selecting features which includes the n-grams, TF- IDF and Word
Embedding techniques. They have used both TF-IDF and n-grams approaches for
ordinary machine learning and TF-IDF (Term-Frequency, Inverse Term Frequency),
Word Embeddings (Word2Vec) and LSTM techniques for deep learning to display
text as a set of numbers. In the last stage of their proposed methodology reviews are
labeled as spam and ham using both traditional machine learning and deep learning
classifiers. SVM, KNN and Naive Bayes (NB) classifiers are used for detecting
fraud reviews instead of frequently used machine learning techniques. They have
used Multilayer Perceptron (MLP), CNN and RNN deep learning approaches for
identifying fraud (they have utilized LSTM, which is a version of RNN).

2.4 Unsupervised Learning Method for Detecting

Fake Review
The use of supervised learning isn’t always appropriate due to the challenge of cre-
ating accurately labeled datasets of fake review . This issue is solved by using
unsupervised learning methods as they don’t require labeled data. For the pur-
pose of identifying untruthful reviews, a novel unsupervised text mining model was
created, incorporated it into a semantic language model and compared it to su-
pervised learning techniques[8]. Their research provides an approximation method

7
for determining the level of authenticity for reviews based on the duplicate identi-
fication findings estimating the similarity of semantic contents among reviews that
used a Semantic Language Model.They created a high-order concept of association
mining in addition to performing unsupervised fake review detection to extrapo-
late idea association knowledge that is context-sensitive. The experimental findings
demonstrate the e↵ectiveness of semantic language modeling and text mining-based
computational models for the identification of fraud reviews, as well as the high
proportion of duplicate fake reviews being found that may be attained using unsu-
pervised approaches.

2.5 Text Reuse Detection

Text reuse is the practice of repeatedly using text from earlier works. Text reuse
detection has been the subject of in-depth research investigations in the context of
online searches.For example the identification of duplicate or near-duplicate doc-
uments [9,10]. The resemblance[9] determines whether two (web) documents are
almost identical, that is, they di↵er only in their alterations of the same information
as in - minor corrections, web-master signature, formatting, logo, capitalization etc.
When the likeness is close to 1, it is likely that the two papers are identical to one
another. The resemblance is a value within 0 and 1 which is specifically detailed
below. They have presented a technique that can exclude almost similar documents
from a library of hundreds of millions of files by computing individually for each
document a vector of characteristics just under 50 bytes long and examining just
these vectors instead of complete documents.

2.6 Feature Engineering for fake review detection

Features are either developed or retrieved from data in feature engineering. Var-
ious elements that may be extracted from reviews have been used in previous re-
search with words from the review’s text being the most common. The bag of words
method often employed for this in which each review’s characteristics are either a
single word or a small set of terms that may be found inside the review’s content.
Less frequently, syntactical and lexical characteristics[11] or features describing the
activities of reviewers have been included in research as extra components of the
items, reviewers and reviews.In general, training a classifier with many types of fea-
tures has led to greater performance than training it with a single type of feature.
According to a research by Mukherjee et al. [12], using the unusual behavioral traits
of reviewers outperformed using the reviews’ language features.

8
Chapter 3

Background Studies

Nowadays, it seems like there are fake reviews everywhere, making it difficult for
shoppers to determine whether goods or companies are reliable. There is always a
chance that the reviews you are reading are fraudulent, whether you are shopping
on Amazon, researching a restaurant on Tripadvisor or reading about a possible
employer on Glassdoor. In this project, we’ll discuss the history of fake review de-
tection, examine some of the features that models use to distinguish between genuine
and fake reviews and opinion spam and develop a basic fake review detection model
that makes use of Tfidf Vectorizer and a number of machine learning algorithms to
distinguish between the two.
The fact that phony reviews can take many di↵erent forms is one of the reasons why
they might be difficult to detect. Of course, there are two basic categories of false
reviews: those produced by humans and those produced by computers. However,
reviews created by people as well as those created by computers can have a positive
or negative tone and can be intended to enhance or lower the overall rating as well
as the number of reviews to give the score more legitimacy[37]. The various false
review kinds you might come across generally fit into one of the four categories
below:

3.1 Computer generated reviews

Fake computer generated reviews can be produced using AI text generation mod-
els. Machine learning may be used to construct and detect these AI reviews as
demonstrated by data science researchers Salminen et al. (2022)[18].

3.2 Human generated via review farms

Through review farms, which promote their services on Facebook and other web-
sites fake reviews can be bought in volume. He, Hollenbeck and Prospero (2022)
researched these phony review companies and discovered that they would be paid
to create evaluations for a range of Amazon products including those with plenty
of reviews and high average ratings. They discovered that businesses’ share of one-
star ratings considerably increased after using phony review services indicating that
review manipulation was particularly common for cheap goods[22].

9
3.3 Human generated fake negative review
Disgruntled customers, ex-employees or rivals that wish to harm a product or com-
pany’s reputation by saturating it with maliciously negative evaluations are known
as human-generated false negative reviewers[22].

3.4 Human generated fake positive review

All review sites that accept them are riddled with human-generated phony positive
evaluations whether they come from restaurants, Amazon marketplace sellers, e-
commerce merchants or HR departments attempting to bury the Glassdoor reviews
of dissatisfied former employees who have criticized the organization[22].
In-depth studies of phony reviews by researchers have uncovered a wide range of
potential characteristics that both people and computer models can use to distin-
guish between a fake and a genuine review. Since false reviews frequently utilize
similar language, especially if they are created by the same individual, business or
review farm, the review text itself is typically the most crucial component[16]. To
identify phony reviews, there are numerous non-text criteria that can be exploited.
In Theodoros Lappas’ paper from the 2012 International Conference on Application
of Natural Language to Information Systems, it examines the many strategies used
to avoid detection and make phony reviews appear legitimate. It is written from
the attacker’s point of view.
Review length sentiment: The review’s word count may reveal if it’s authentic or
not.
Sentiment: False reviews frequently have more extreme sentiments, either being
positive or extremely negative.
Helpfulness: There may be a link between phony reviews and lower helpfulness
scores when review helpfulness is a platform indicator.
Reviews per user: The user who left the review might have just signed up because
they made a false account with the express goal of leaving a fraudulent review be-
cause some reviews may be manufactured by bots. Models may find it advantageous
to be able to detect the average amount of reviews per user.
Verified reviews: Some review sites, like Trustpilot don’t demand proof that you’ve
used a company’s goods or services before leaving a review, making them vulnerable
to fraudulent reviews. When a review is ”verified,” it means that it was written
after the retailer requested it in response to a tracked customer purchase.
Stealth: ”Stealth evaluates the capacity of the review to blend in with the cor-
pus,” claims Lappas (2012). Fake reviews may aim to appear like other evaluations
from the company because they may stand out if they are written in a completely
distinctive manner from the others.
Coherence: We may occasionally come across reviews that give a product or service
a very low rating but yet include text praising it. It ”evaluates whether the assigned
grade is in line with the viewpoints conveyed in the review’s content,” according to
Lappas (2012), who labels this ”coherence.”
Readability: Readability has been found to be a useful tool for identifying some
bogus reviews by a number of machine learning researchers. For instance, Flesch
Reading Ease (FRE), which other writers have also included in their models was
employed by Lappas (2012).

10
Review text: The one element that phony review detection models use the most is
the review text. Text is converted into a Bag of Words using NLP techniques like
count vectorization and TF-IDF and then Naive Bayes text-classification algorithms
are employed to determine whether or not the review is fake.

3.5 Fake Reviews Detector: Broad Overview

Deception variables yields accuracy levels of about 90%. The dataset that was used
for the same is likewise open to the public. Additionally, ordinary n-gram text cate-
gorization methods do significantly better than human judges at detecting negative
false opinion spam[37]. Multi-layer perceptrons were applied for classification in ar-
tificial neural networks’ encouraging text classification performance. By examining
user activity attempted to identify spamming networks using the frequency with
which reviewers posted for the same items.

3.5.1 Natural Language Processing techniques

Text mining and natural language processing are terms used to describe the under-
standing and analysis of natural language using computer methods and algorithms.
It is a prominent topic of research in the field of applications of artificial intelligence.
Text mining and natural language processing studies have been studied since the in-
vention of computers. Thanks to continuous, detailed research on machine learning
and data mining algorithms, existing text mining technologies have achieved suc-
cess in automated abstraction, automatic question answering, web relational network
analysis and anaphora resolution[22]. An interdisciplinary field known as bioinfor-
matics was born out of the success and advancement of the Human Genome Project.
It makes predictions and resolves actual genetics-related scientific problems utiliz-
ing computer and statistical informatics. The three main steps in bioinformatics
are data storage, retrieval and analysis. The National Center for Biotechnology In-
formation developed the Online Mendelian Inheritance in Man database, the Gene
Expression Omnibus database, the Sequence databases for storing DNA and protein
data, the PubMed database for preserving biological and medical literature among
other databases.[23].

3.5.2 Machine Learning in NLP

Artificial intelligence and machine learning have grown in popularity over the past
several years. Their methodologies and ideas are now used in a vast area of products
and most applications and appliances require them. The automatic identification
of essential emails and speedy responses in Gmail are examples of applying ma-
chine learning. Today, we can declare with confidence that artificial intelligence
and machine learning can extricate a person from many technological processes[24].
Machine learning is the scientific study of the statistical methods and algorithms
that computer systems use to successfully do a certain task without using explicit
instructions and instead relying on patterns and inference. It is regarded as a com-
ponent of artificial intelligence. Machine learning algorithms build a mathematical
model using sample data or ”training data” without being explicitly instructed to

11
do so in order to make predictions or judgements[24, 25]. Machine learning algo-
rithms come in five di↵erent flavors: supervised, semi-supervised, active learning,
reinforcement learning and unsupervised learning. A branch of computer science, in-
formation engineering and artificial intelligence called ”natural language processing”
is interested in how computers interact with human (natural) languages, particu-
larly how to teach computers to process and analyze massive amounts of natural
language data.

3.5.3 Support Vector Machine (SVM) Machine Learning

Model
Support-vector machine (SVM), also known as Support - Vector networks in machine
learning are supervised learning models with corresponding learning algorithms that
examine data used for regression and classification analysis. A well-liked machine
learning method for solving classification and regression issues is the Support Vec-
tor Machine (SVM) algorithm[26]. A separating hyperplane serves as the formal
definition of a Support Vector Machine (SVM), a discriminative classifier. In other
words, the method generates an optimum hyperplane that classifies fresh samples
given labeled training data (supervised learning). This hyperplane, which divides a
plane into two portions in two-dimensional space, has one class on either side.

3.5.4 Logistic Regression Machine Learning Model

Logistic regression is most popular Machine Learning methods for two-class catego-
rization. A statistical approach is used to forecast binary classes. - Applications like
machine learning and data mining need the use of classification techniques. Used
for several classification-related concerns, such as diabetes prediction and spam de-
tection, among others[28].

3.5.5 Term Frequency Inverse Document Machine Learning

Model
The preferred method for presenting documents as feature vectors is as described
above. The abbreviation TF-IDF means ”Term Frequency, Inverse Document Fre-
quency.” According to a document and the full corpus, a word’s importance is mea-
sured using TF-IDF[30]. Authentic Review: ”Lovely, bright, and tidy room. Excel-
lent views, peaceful, a decent bed, etc. Excellent sta↵ and overall top-notch service.
For the money, the quality and value were superb. Also at a great location. False
rating: ”Dark and disorganized room. Loud, uncomfortable beds, poor views, etc.
Poor overall service and unfriendly sta↵. It was of very low value and quality, and
the location was awful. The review sample stated above amply demonstrates how
spammers reverse the polarity of aspects like room, bed, sta↵, and location. Tags
like stunning, clean, and bright might be used to infer the atmosphere of the place.
The polarity of the room is altered through the use of phony evaluational words like
untidy and dark. This sample was picked from many line evaluations in order to
better demonstrate our technique. The aforementioned example demonstrates that
it is unnecessary to focus on in-depth text analyses of several lines and words be-
cause doing so will just make the computing process take longer[15,16]. The review

12
sample stated above amply demonstrates how spammers reverse the polarity of as-
pects like room, bed, sta↵, and location. Tags like stunning, clean, and bright might
be used to infer the atmosphere of the place. The polarity of the room is altered
through the use of phony evaluational words like untidy and dark. This sample was
picked from many line evaluations in order to better demonstrate our technique. The
aforementioned illustration demonstrates that it is unnecessary to focus on compre-
hensive text evaluations of several lines and words because doing so will just make
the computation process take longer[15, 16]. Because spammers are unfamiliar with
the product, they change a few words to create fictitious reviews. The authors have
discovered that the analysis of fake reviews can be done using part of speech (POS)
tagging. Our approach uses POS tagging to tag nouns and noun phrases as aspects,
adjectives, adverbs, and verbs as sentiments. Deep learning algorithms outperform
traditional classifiers when assessing spam comments. The reason for this is that
traditional classifiers lose their ability to provide precision once they achieve a cer-
tain level of accuracy. Their suggested approach employs a hybrid CNN and LSTM
model for aspect replication and sentiment learning. Extracted aspects and their
respective polarities are fed into a CNN model in order to find aspect repetition.
Then, for training and performance evaluation, the filtered aspect replication is fed
into an LSTM[15].The following are the primary contributions of this research work:
1. Aspect and polarity from reviews are extracted using eective POS tagging-based
algorithms. 2. Their method is superior to existing approaches because it computes
aspect replication in spam reviews rather than whole text replication. 3. In order
to analyze bogus reviews, extracted characteristics and feelings are fed into deep
learning models (RNN and LSTM). 4. Ott and Yelp filter datasets are used fre-
quently in experiments to analyze accuracy. Analysis of the results of experiments
demonstrates that our suggested strategy oers greater precision and accuracy when
compared to current approaches. Aspect extraction and spam opinion approaches
that are related to this study are reviewed. Background information and prelim-
inary steps are provided for aspect extraction methods, RNN, LSTM architecture
and spam opinion detection. The suggested strategy is explained. The design and
analysis of experiments are discussed[16]. To identify fraudulent reviews, previous
research has concentrated on the reviewer’s behavior or the review wording. The
most typical technique employed by spammers to produce phony reviews is the us-
age of sentiments and material that has been taken directly from reviews. This part
surveys current state-of-the-art methods to evaluate the benefits and drawbacks of
previous research, which inspires us to suggest a practical method for identifying
false reviews with increased precision and reduced computing time.

3.6 Analysis of fake reviews using review text

The use of neural networks to identify false opinions using document-level represen-
tations is investigated. The authors of this work state that earlier research only paid
attention to discrete qualities that were based on linguistic viewpoints. Discourse
and document-level semantics are taken into consideration in this research project.
In order to analyze the discourse semantics, a recurrent neural network receives in-
put from the CNN model, which is used to learn representation from phrases. Three
datasets were used for the experiments and it was determined that the proposed
strategy enhanced accuracy. AutoEncoder (DAE) and Vector Distributed Bag of

13
Words (PVBOW) is suggested. The pre-processing of the dataset uses lemmatiza-
tion and tokenization. For hidden layers, ReLU is utilized as the activation function,
and Sigmoid function is used for output layers. For the analysis of the experiment,
gold standard dataset is employed. When evaluating the suggested strategy, accu-
racy, F1-measure, precision and recall are taken into consideration. Reviewers are
given a reputation score to determine whether they are legitimate or spam. This
method has the benefit of not requiring huge instances to be labeled[18]. The ap-
plication of additional k-centre clustering depends on time interval. Reviews posted
by spammers allegedly have significant emotional inclinations. Combining content
attributes with reviewer behaviors yields reputation value. Amazon music product
reviews serve as the dataset for the experiment analysis. In this study, precision,
recall and F-measure are used. In comparison to current methodologies, the perfor-
mance of the proposed methodology is better. For experiment analysis in this paper,
a dataset of Amazon product reviews is employed, together with the entire review
content. To extract aspects, a deep convolution neural network is used. Seven-layer
neural networks have been utilized by the authors to increase aspect extraction’s
precision. It is claimed that linguistic patterns and conditional random fields have
shortcomings that need to be fixed. For experiment analysis, datasets from Google,
Amazon and SemEval are used. The proposed technique outperforms LP when pre-
cision values and language patterns are examined. There are several neural network
designs that are explicitly discussed for spam opinions. This study work also makes
the claim that conventional machine learning algorithms do not o↵er the semantic
information of reviews that is required for the analysis of misleading spam opinions.
Numerous neural network topologies including CNN, RNN, LSTM and GRU are
used in experiments. In terms of accuracy, CNN outperforms other models. It’s
because CNN can extract sophisticated and intricate features from user opinions.
Information from the review text is extracted using the N-Gram model. The en-
semble method which combines CNN and N-gram is applied. On the Yelp dataset,
experiments are run to evaluate the e↵ectiveness of the suggested approach. The
drawback of this method is that it relies on a number of reviewer characteristics and
reviewer behavior to identify phony reviews. The computational complexity rises
as a result. For the purpose of detecting false reviews, various machine learning
and deep learning models are employed. On the Yelp and Ott datasets, the per-
formance of CNN, LSTM, SVM and kNN models is assessed. The drawbacks of
this strategy include the usage of a single model for the detection of false reviews
and the possibility of word embedding and hyperparameter settings enhancement.
Utilizing semi-supervised learning, bogus reviews can be found. The authors claim
that while precise labels and extensive data are needed, labeled datasets can present
a problem. In order to choose features that can perform better in less computing
time, this approach needs optimization[19].

3.7 Analysis of fake reviews using reviewer be-

havior
Entities are suggested as a method for detecting opinion spam. In this method,
entities are given weights based on how important they are. This study discusses
spam characteristics including review polarity intensity, words, rating deviation, re-

14
views per day, etc. Utilizing the evaluation criteria Accuracy, Precision, Recall, and
F-measure, the suggested approach is verified. This study uses a range of char-
acteristics, including content-based, behavior-based, relation-based, and proposed
features, to measure correctness. Accuracy is improved by using the indicated set
of qualities. The authors have proposed a method based on singleton reviews. Few
research, according to the authors, have focused on reviewers who have only pub-
lished one review. It is possible for reviewers to submit spam reviews by changing
their usernames. Between reviews, textual and semantic similarity is computed. On
the datasets from Yelp and Trustpilot, experiment analysis is done. By utilizing
the suggested approach, precision and F1 score are increased. To identify spam
opinions, autoencoder and neural random forest are utilized. These models were
chosen because random forest combines many decision trees while autoencoder can
employ unsupervised representations in features. This study took into account both
reviewer behavior and review substance. Features include things like the summary’s
word count, the ratings’ entropy, when they were given, etc. The experiment analy-
sis uses the Amazon review dataset, and the evaluation metrics used are precision,
recall, and F1-measure. When compared to the current approach, it has been shown
that the proposed approach o↵ers more accuracy. The computational challenge is in-
creased by the use of the complete review text for semantic analysis. From content-
and behavior-based aspects, 133 traits are gathered. The class imbalance in the
datasets is mentioned. To resolve this, a random sampling is employed. Accuracy
is improved by sampling[20]. This approach does not, however, use many classifiers
and there is room for development to reach satisfactory accuracy on sizable datasets.
Reviewer behavior is examined in order to identify bogus reviews. Star User, De-
viation Rate, Bias Rate, Review Similarity Rate, Review Relevancy Rate, Content
Length, and Illustration are a few of these. In order to improve the classification,
sophisticated neural network-based models must be used. The novelty, advantages,
disadvantages, assessment criteria and dataset of contemporary systems to fake re-
view identification are analyzed. Our analysis reveals that whole text reviews are
employed in previous studies which raises the computational di culty. Reducing
computational complexity is necessary. Additionally, shallow architectures are em-
ployed for the detection of false reviews, necessitating the optimization of the neural
network design through the use of dropout, e cient feature selection and hyperparam-
eter tuning. Aspects are not given significant attention in current study activities.
Only pertinent features are used in our suggested method, which reduces computing
complexity. Iterations of the complexity analysis are made. Additionally, the major-
ity of contemporary approaches use deep learning or conventional machine learning
models. In our suggested method, a hybrid deep learning model is used to take
use of the benefits of CNN and LSTM as mentioned. Recent research has shown
a significant increase in interest in the use of NLP for opinion spam identification
and false review detection. Due to this, a large number of literature reviews and
research articles have been written with the specific goal of identifying fraudulent
reviews or opinion spam. The emphasis of this section is on showcasing earlier re-
search in this area. Classification models may help distinguish between reviews and
classify them as either real or fake, which makes them quite useful when it comes
to applying machine learning to identify fraudulent reviews. A dataset of 2.14 mil-
lion reviews, many of which were duplicates, was utilized to generate classification
predictions for reviews using the conventional machine learning technique of logistic

15
regression. A rudimentary framework for spotting spam reviews was given using
a combination of supervised, semi-supervised, and unsupervised learning. For the
purpose of detecting fraudulent reviews, a thorough approach that included data
gathering, data pre-processing, and machine learning was o↵ered. Semantic simi-
larity between terms at the review level and topic modeling were employed in an
e↵ort to spot fake reviews that had been written by the same person but published
under various names. The popular Yelp algorithm for identifying fake restaurant
reviews performed well for behavioral features but not so well for linguistic features,
it was also found. Singular Vector Machines (SVM) have been shown to perform
better in text classification than other traditional data mining algorithms[19]. It is
also possible to tell whether a review is phony or real by counting how many users
have given it a helpful vote or upvote, which measures how beneficial the review
is. Additionally, it was discovered that combining a classifier with both n-gram and
psychological over short periods of time. They also took into account the frequency
with which other users posted for the same products during those same periods of
time. For topic modeling, Probabilistic Latent Semantic Analysis (pLSA) can be
used to automatically extract aspects and organize them into dierent categories. La-
tent Dirichlet Allocation (LDA), a generative probabilistic model that assumes that
each document is built of a variety of themes and that each topic is a probability
distribution over words, can also be used for topic modeling, which was introduced
as a technique for identifying phony reviews. For text mining, three additional fea-
ture categories were introduced: review density, semantic similarity, and emotion.
The review centric approach and the reviewer centric strategy were presented as an
overview of the current detection techniques.

16
Chapter 4

Methodology

4.1 Workflow
In our little time of research, we believe that the workflow is the most crucial com-
ponent of a study. Our work cycle is outlined in Figure 4.1. We used this approach
in order to perform as intended or to maximize the outcome. The initial step in
our approach was choosing the data sets. The dataset was then preprocessed. In
data preprocessing we removed the punctuations and stopwords. Then, we replace
the target value. Then we split the data where we split the value as train, test and
val. Then we jump in the next part word embedding. In word embedding part we
use tfidf and glove. Then we will use Support Vector Machine (SVM) and Logistic
Regression (LR) in Glove. Moreover, to do the train test part we will use the value
of glove from Logistic Regression and Support Vector Machine Model. By this way
we get the accuracy. By using this accuracy we will find do the performance train.

Figure 4.1: Workflow

17
4.2 Dataset
The dataset we used in this study has been taken from OSF — Pdfs from dot .
This dataset contains 40,431 customer reviews of di↵erent products. Here we have
2 di↵erent types of reviews as in computer generated reviews and original reviews
from people. Computer generated reviews are labeled as CG and original reviews
from people are labeled as OR in the dataset.
We faced difficulty while searching for our desired dataset. At first we chose a
dataset from Amazon which contained approximately 14 lakhs reviews which was
quite large for our research. Most of the datasets available online are more or
less of the same size and quite large and robust. For this reason, we had to surf
the web well to finally find our desired dataset. A proper dataset is needed because
using raw data companies may establish baselines, benchmarks and goals in order to
advance. Data enables measurement allowing you to establish baselines. A baseline
is the condition of a place before a particular remedy is applied.

4.3 Preprocessing
In order for businesses to undertake data mining, evaluate the data and process it
for business activities, raw data must be transformed into legible and defined sets.
Preprocessing by deleting missing or inconsistent data values caused by human or
machine mistake data may improve the accuracy and quality of a dataset making it
more dependable. Preprocessing raw data e↵ectively can boost its precision, which
will boost projects’ quality and dependability. Data duplication is a possibility
while collecting data, and removing them during preprocessing can guarantee that
the data values for analysis are consistent, assisting in the production of reliable
results. At first we splitted the dataset into two groups where 90% of the dataset
was reserved for training the model and 10% was kept for testing. Then again we
splitted the training dataset into two groups where 90% of the dataset was assigned
for training the model and 10% was kept for validation. As a result, we got 81% of
the main dataset as training set, 10% as testing set and 9% as validation set. We
kept 9% of the dataset as validation for tuning the dataset.
Data is transformed into discrete variables as the initial step in dataset pre-processing.
To make things function, we had to turn the attribute domains into discrete vari-
ables. Machine learning models cannot utilize text on their own. They anticipate
receiving numerical input.
Data engineering and feature engineering are both involved in the preprocessing of
the data for ML. The prepared data is subsequently tuned through feature engineer-
ing to provide the features that the ML model anticipates.
Here we converted every text of our dataset to lowercase as this dataset will be
used as our main input. For output we will need OR and CG labeled data but as
machine learning models cannot use text on their own we replaced ‘OR’ as 0 and
‘CG’ as 1. For this, now we will get output as 0 or 1. Then we removed all the
punctuation marks from our dataset and created a new column named twp. Then
from this column we removed all the stop words(e.g “a”, “the”, “is”, “are” and etc.)
and created another column named tws. For further operation we used this tws
column.

18
4.4 Model Implementation
We examined several supervised classification algorithms for this work and Support
Vector Machines (SVM), a well-known technique that has been successful in many
areas, including deception detection which produced the best results. Its ability to
cope with things that would not be linearly separable in the feature space, using
kernel functions that move the entities in a higher dimension space, where the linear
separation is achievable is what determines how successful it will be. As a result,
the kernel function selection is essential to the models’ efficacy. Although texts are
typically represented in vector space by sparse vectors, radial kernels performed the
best in our trials even though linear kernels are typically thought to be e↵ective
for text categorization. Because of the relatively modest size of our corpus, which
would have prohibited the training of accurate word embeddings relevant to the
task, we chose to employ SVMs rather than the Deep Learning techniques that have
been frequently used in recent literature on text categorization. The majority of the
literature on deception detection uses more conventional classifiers, which allows
for a more direct comparison of our results to other researchers’ findings. Tenfold
cross-validation was used as the validation approach during the model’s training.
This paper’s evaluation of the widespread practice of generating synthetic data sets
for the identification of false reviews via crowdsourcing was one of its primary objec-
tives. This approach is frequently employed in the literature, which typically reports
positive outcomes. Our findings, however, indicate that this strategy might not be
the best one. Additionally, we observe a significant change in the prediction accu-
racy depending on how close the training and test sets are in terms of the domain.
For model implementation, first of all we built the model and installed the pack-
ages. Then, we imported the packages and loaded all the data. If there is any data
imbalance then checked the data imbalance. After preparing the data, we created
features from punctuation, tokenized the data and used stopword removal. Then
after applying porter stemming and rejoining words we created training testing data
and run the model selection process. By this we accessed our proposed model.

4.5 TF-IDF
The metric TF-IDF stands for frequency-inverse document. Frequency can be used
to assess how important or pertinent string representations (words, sentences, lem-
mas, etc.) are inside a document in comparison to other texts. It is utilized in the
fields of machine learning and information retrieval (IR). In TF-IDF there are two
parts: TF (Term frequency) and IDF (Inverse document frequency).
The TF-IDF states that a term’s value is inversely correlated with how frequently it
appears in documents. While TF o↵ers information about how frequently a phrase
appears in a document, IDF provides information about the relative rarity of a term
in the collection of documents. By averaging these figures, we might arrive at our
ultimate TF-IDF value. The following mathematical formulas are used to determine
the TF-IDF score for the word t in the document d from the document collection
D:

tf idf (t, d, D) = tf (t, d).idf (t, D) (4.1)

19
In Tfidf vectorizer we applied fit transform on our trained dataset and applied only
transform on validation and test dataset.

4.6 GloVe Vectorizer

Global vectors for word representation is referred to as GloVe. The global word-
word co-occurrence matrix from a corpus is combined with this Stanford-developed
unsupervised learning technique to create word embeddings. The resulting embed-
dings in vector space emphasize the word’s intriguing linear substructures. In our
proposed framework, we took the txt file of GloVe Vectorizer from Stanford website.
For every word it will have 100 floating points and the file had approximately 4 lacs
floating points for di↵erent words. We took this into our dataset then we converted
it to dictionary. Then we calculated floating point value for each word of every
review from our dataset and stored them to a list. We added all the floating point
value of every word of a single review and divided it by the total number of words
and stored it in our dataset. After that we use this in SVM and LR as input. But
for LSTM we did not add the floating point of every word for a sentence. Here we
took floating value for every word and append the value in a 3d list. In that 3D list
every row represents a review. Then we converted that 3d list into dataset and use
that dataset for input in LSTM model.

4.7 Support Vector Machine Algorithm(SVM)

Currently the majority of machine learning tasks involve classifying images, trans-
lating languages, managing vast quantities of sensor data and making predictions
about the future value based on the present.For both classification and regression
issues, Support Vector Machine (SVM), amongst the most well-liked supervised
learning techniques, have been used. The SVM method seeks to define the best line
or decision boundary which can divide n-dimensional spaces into classes in order
to quickly categorize new data points in the future. This boundary’s best option
is known as a hyperplane.As they select the decision boundary that optimizes the
distance from the nearest data points of all the classes, SVMs vary from other clas-
sification techniques. How SVM functions is discussed briefly below -
i) Linear Case - The problem of two classes with N training samples should now be
taken into consideration. A Support Vector (SV) Xi comprised of various ”bands”
with n dimensions is used to describe each sample. Yi is written on a sample’s label.
In the event of two classes, we take into account the labels 1 for the first class and
+1 for the second. The function is defined by the SVM classifier -

f (x) = sign(<!, X> + b) (4.2)

which determines the ideal separating hyperplane , where ! is the hyperplane’s

|b|
normal and ||!|| is the hyperplane’s distance from the origin perpendicularly. The
label of the sample is provided by the sign of f(x). The SVM seeks to maximize the
distance between the support vector and the ideal hyperplane. Thus, we search the
min |!|
2

20
Using the Lagrange multiplier makes this process simpler. The issue has to be
resolved:
Ns
X
f (x) = sign( .yi .↵i <x.xi > + b) (4.3)
i=1

where ↵i is the Lagrange multiplier

ii) Nonlinear Case -
In our proposed framework, we used four kernels (rbf, linear, poly, sigmoid). Data
can be entered and then changed into the format required for processing using a
kernel function. The word ”kernel” is used because a Support Vector Machine’s
window for data manipulation is given by a series of mathematical operations.
Three kernels are commonly used: The polynomial kernel:

K(x, xi ) = (<x, xi > + 1)p (4.4)

The sigmoid kernel:

K(x, xi ) = tanh(<x, xi > + 1) (4.5)

The RBF kernel:

|x xi |2
K(x, xi ) = exp 2 2 (4.6)

We got di↵erent accuracy scores for each of them in di↵erent regularization param-
eter(CS = 1.0,1.1,1.2). We got the highest accuracy score for rbf which is 89.86%.

4.8 Logistic Regression(LR)

Another type of supervised learning technique is logistic regression. Binary logistic
regression is the most often used application of logistic regression where the outcome
is a binary decision (yes or no). A logistic regression model forecasts the value of
P(Y=1) as a function of X mathematically.It may be applied to many categorization
issues, such as spam detection, diabetes prediction, cancer diagnosis, etc.
Here is an illustration of a logistic regression formula:
bo +b1⇤x
y = e 1+e( bo +b1⇤x) (4.7)

If we take b0 is the bias/intercept term, then b1 = coefficient of the single input value
(x) and y = the anticipated output. You must learn the associated b coefficients
(constant real values) for each column in your input data from your training set.
In our proposed framework we used di↵erent solvers(newton-cg, lbfgs, liblinear, sag,
saga).We got di↵erent accuracy scores for each of them in di↵erent regularization
parameter(CS = 1.0,1.1,1.2). Each observation gives an equal amount of informa-
tion, which is a key premise of logistic regression. There is a chance to give a Weight
variable while using Analytic Solver Data Mining. The user can assign a weight to
each entry by using the Weight variable. We got the highest accuracy score for
newton-cg and sag in this case which is 87.94%.

21
4.9 BiDirectional LSTM
Even though we got a decent result with Logistic Regression and Support Vector
Machine, we aimed to improve the accuracy of our results by utilizing the power
of neural networks. We chose to work with Recurrent Neural Networks (RNNs) as
they have a proven track record in working with sequential data. However, due to
the long-term dependency problems of traditional RNNs, we decided to implement
Long Short-Term Memory (LSTM) networks.
Initially, a single-directional LSTM with Glove Vectorizers provided us with an ac-
curacy of 85%. However, this performance was not as high as we had hoped. There-
fore, we attempted to increase the accuracy further by implementing a Bidirectional
LSTM model. The model consisted of one Global Max pooling 1-dimensional layer,
followed by a batch normalization layer to speed up the training process with higher
learning rates. Additionally, there were three back-to-back dropout layers, followed
by one dense layer each, with the first two having relu and the last having sigmoid
as the activation function. The dropout layers were crucial in preventing overfit-
ting by shutting down the contributions of certain neurons. The model had an input
dimension of (237 x 100). To compile the model, we used Root Mean Squared Propa-
gation (RMSProp) optimizer to accelerate the Gradient Descent process. Compared
to AdaGrad, RMSProp performed faster as it utilizes the decaying average of the
partial derivatives. The loss was calculated using the Binary Cross-Entropy loss
function, as it is best suited for datasets with 2 target classes.

Figure 4.2: Bidrectional LSTM

4.10 Confusion matrix, Accuracy, F-1 Score , Re-

call and Prescision
A confusion matrix is a table that shows the results of a classification problem,
including the predicted and actual values of a classifier. It is used to visualize the
outcomes of the prediction and helps to understand the performance of the classifier.
There are 4 types of values in confusion matrix.

1)T rueP ositive (4.8)

2)T rueN egative (4.9)

22
 3)F alseP ositive (4.10)

4)F alseN egative (4.11)

Figure 4.3: Confusion matrix

A measure of accuracy means the proportion of correctly classified data cases to

the total number of data cases, which indicates how many data cases were correctly
classified.
TN + TP
Accuracy = (4.12)
TN + FP + TP + FN
The precision of a classifier’s accurate positive predictions is measured. It is defined
as the proportion of accurate positive predictions to all of the classifier’s positive
predictions.
TP
P recision = (4.13)
TP + FP
Recall is a metric for how many out of all the positive cases in the data that the
classifier correctly predicted.
TP
Recall = (4.14)
TP + FN
The F1-Score is a statistic that combines a classifier’s precision and recall. It is
computed as the harmonic mean of recall and precision, which is a separate method
than calculating the variables’ ”average.” The regular arithmetic mean is seen to be
less suitable for ratios like precision and memory than the harmonic mean.
P recision ⇤ Recall
F 1Score = 2 ⇤ (4.15)
P recision + Recall

23
Chapter 5

Experiments and Results

5.1 Result analysis

The dataset for this study was collected from OSF — Pdfs from dot. 40,431 con-
sumer reviews of di↵erent products are included in this dataset. Here, there are
two di↵erent kinds of reviews: those produced by computers and those that real
people have written. In the dataset, computer produced reviews are denoted as CG
and manually authored reviews as OR. There are total 20216 reviews which is com-
puter generated and other 20215 reviews which is real reviews from people. In our
datasets there we used two columns one is text and another one is label where text
column represents all the reviews and label represents if the review is CG(Computer
Generated) or OR(Original Review). At first, we splitted our dataset in two set one
is for train set where the test size is 10% of the main dataset and train set is 90%
then again we splitted the train set into train set and validation set where train
set has 81% data of the main dataset and validation set has 9% data. We splitted
twice one is for train test another one is for train validation. The second split we
did for tuning our model. Model tuning is a process to discover the optimal values
in order to maximize the performance of the model. So after that we implemented
TF-IDF vectorizer on input part which are x-train,x-test and x-val. Next we also
implemented GloVe vectorizer on our input part of the dataset. Here we have used
Support Vector Machine(SVM). Logistic Regression(LR) and LSTM. We used SVM
and LR two times one with TF-IDF vectorizer value and other one with GloVe
vectorizer value.

5.1.1 SVM and LR Results using TFIDF:

First we used a loop to try di↵erent types of parameters in our SVM model. In
parameters we set four types of kernel ( ‘rbf’, ‘linear’, ‘poly’, ’sigmoid’) then we
set max iteration -1 and we used di↵erent regularization parameter( CS = 1.0, 1.1,
1.2). We checked accuracy with each CS for each kernel. We got the best accuracy
(89.50%) and weighted f1 score(0.84) when kernel = “rbf” and c=1.2 and we got
less accuracy(84.36%) and weighted f1 score(0.84) when kernel = “poly” and CS=
1.0. . Then we used train set and test set in SVM model by setting the value of
kernel = “rbf” and c=1.2 for these values we got weighted f1 score 0.90 and accuracy
90.31%. Then we did the same thing using LR but here instead of using kernels
we used solvers(newton-cg, lbfgs, liblinear, sag, saga). Then we set max iteration

24
500 and we used di↵erent regularization parameter( CS = 1.0, 1.1, 1.2) here also.
We checked accuracy with each CS for each solver. Then we used train set and
validation set in LR model by setting di↵erent values of solver and CS. For solver =
“newton-cg” and CS=1.2 we got weighted f1 score 0.88 and accuracy 87.94%. We
got the best accuracy in this case. We got less accuracy(87.80%) and weighted f1
score(0.88) when kernel = “sag” and CS= 1.0. Then we used train set and test set
in LR model by setting the value of For solver = “newton-cg” and CS=1.2 for these
values we got weighted f1 score 0.89 and accuracy 88.92%.

Model Precision Recall f1-Score Accuracy

RBF 1.0 0.92 0.87 0.89 0.90
RBF 1.1 0.91 0.88 0.89 0.90
RBF 1.2 0.79 0.88 0.89 0.90
linear 1.0 0.89 0.87 0.88 0.88
linear 1.1 0.89 0.87 0.88 0.88
linear 1.2 0.89 0.87 0.88 0.88
poly 1.0 0.94 0.74 0.83 0.84
poly 1.1 0.94 0.75 0.83 0.85
poly 1.2 0.94 0.76 0.84 0.85
sigmoid 1.0 0.89 0.85 0.87 0.87
sigmoid 1.1 0.88 0.85 0.87 0.87
sigmoid 1.2 0.88 0.85 0.87 0.87

Table 5.1: Results of using di↵erent parameters of SVM with TFIDF

Model Precision Recall f1-score Accuracy

newton-cg 1.0 0.90 0.85 0.88 0.87
newton-cg 1.1 0.90 0.86 0.88 0.87
newton-cg 1.2 0.90 0.86 0.88 0.87
Lbfgs 1.0 0.90 0.85 0.88 0.87
Lbfgs 1.1 0.90 0.86 0.88 0.87
Lbfgs 1.2 0.90 0.86 0.88 0.87
Liblinear 1.0 0.90 0.85 0.88 0.87
Liblinear 1.1 0.90 0.86 0.88 0.87
Liblinear 1.2 0.90 0.85 0.88 0.87
Sag 1.0 0.90 0.86 0.88 0.87
Sag 1.1 0.90 0.85 0.88 0.87
Sag 1.2 0.90 0.86 0.88 0.87
Saga 1.0 0.90 0.85 0.88 0.87
Saga 1.1 0.90 0.86 0.88 0.87
Saga 1.2 0.90 0.85 0.88 0.87

Table 5.2: Results of using di↵erent parameters of LR with TFIDF

25
 Figure 5.1: Confusion Figure 5.2: Confusion
matrix for SVM usign matrix for LR usign
GloVe GloVe

5.1.2 SVM and LR Results using GloVe:

Secondly for GloVe vectorizer we got best accuracy=79.44% and weighted f1 score=
0.79 using SVM. We got less accuracy 45.07% and weighted f1 score= 0.45 when
kernel = “sigmoid” and CS= 1.1. Then for train and test dataset we used kernel
= “rbf” and CS = 1.2 as we got best accuracy for these values. After applying it
we got 0.8 as f1 score and 80.34% as accuracy. Then we did the same thing for LR.
We got best accuracy=76.17% and weighted f1 score= 0.76 using LR when solver
= “lbfgs” and CS=1.0. We got less accuracy 76.12% and weighted f1 score= 0.76
when solver = “newton-cg” and CS= 1.1. Then for train and test dataset we used
solver = “lbfgs” and CS = 1.0. After applying it we got 0.77 as f1 score and 77%
as accuracy.

Model Precision Recall f1-Score Accuracy

RBF 1.0 0.79 0.79 0.79 0.79
RBF 1.1 0.78 0.81 0.80 0.79
RBF 1.2 0.79 0.82 0.80 0.79
linear 1.0 0.77 0.76 0.76 0.76
linear 1.1 0.77 0.76 0.76 0.76
linear 1.2 0.77 0.76 0.76 0.76
poly 1.0 0.80 0.77 0.78 0.79
poly 1.1 0.80 0.77 0.78 0.79
poly 1.2 0.80 0.77 0.78 0.79
sigmoid 1.0 0.46 0.44 0.45 0.45
sigmoid 1.1 0.45 0.44 0.45 0.45
sigmoid 1.2 0.45 0.44 0.45 0.45

Table 5.3: Results of using di↵erent parameters of SVM with GloVe

5.1.3 Bidirectional LSTM accuracy and Loss function:

But for LSTM after 7 epochs, the model achieved an accuracy of 0.9166 and a
binary cross-entropy loss of 0.1533. This was the best score we found among all
the models we tested, although we believe that it can be further improved with
additional tuning and training.

26
 Model Precision Recall f1-score Accuracy
newton-cg 1.0 0.77 0.76 0.76 0.76
newton-cg 1.1 0.77 0.76 0.76 0.76
newton-cg1.2 0.77 0.76 0.76 0.76
Lbfgs 1.0 0.77 0.76 0.76 0.76
Lbfgs 1.0 0.77 0.76 0.76 0.76
Lbfgs 1.1 0.77 0.76 0.76 0.76
Lbfgs 1.2 0.77 0.76 0.76 0.76
Liblinear 1.0 0.77 0.76 0.76 0.76
Liblinear 1.1 0.77 0.76 0.76 0.76
Liblinear 1.2 0.77 0.76 0.76 0.76
Sag 1.0 0.77 0.76 0.76 0.76
Sag 1.1 0.77 0.76 0.76 0.76
Sag 1.2 0.76 0.76 0.76 0.76
Saga 1.0 0.77 0.76 0.76 0.76
Saga 1.1 0.77 0.76 0.76 0.76
Saga 1.2 0.77 0.76 0.76 0.76

Table 5.4: Results of using di↵erent parameters of LR with GloVe

Figure 5.3: Confusion Figure 5.4: Confusion

matrix for SVM usign matrix for LR usign
GloVe GloVe

Figure 5.5: Bidirectional LSTM Accuracy and Loss Function

5.1.4 Discussion:
Word embeddings may be a preferable option for most tasks where tf-idf is employed,
especially where the work may benefit from the semantic similarity recorded by word
embeddings, even if tf-idf is a straightforward scoring method and that is its main
advantage. If GloVe is trained properly it captures all the meaning. In NLP task
GloVe should carry more information than Tf-Idf. Because it gives us 100 value for

27
a word whereas tfidf assign only one value based on how important the word is for
this document. Here,Glove didn’t perform well because here we took Glove value
for each word and add that value with others words of the review and then divided
that value with the number of words that review has so we got an average value
for a review. So SVM and LR might not catch the proper info while training and
that caused for the less accuracy. Besides, we got better accuracy for Bidirectional
LSTM because here we used GloVe value for each word and directly pass the value
to the model. From our analysis we got accuracy 90.30 percent for SVM and 88.92
percent for Logistic Regression where we used TF-IDF and For Glove vectorizer
we got accuracy 80.34 percent for SVM and 77.00 percent for Logistic Regression.
In our proposed architecture we got more accuracy using Tf-Idf than GloVe for
svm and LR. One of the reasons why this happened maybe GloVe was not able to
capture more information than Tf-Idf which could have resulted in underfitting the
model. As the input vector size in Tf-Idf was bigger than GloVe, it might have been
underfitted because of having less features than Tf-Idf. But for LSTM GLove result
did better result.

Number Model Accuracy

1 Bidirectional LSTM 0.9166
2 SVM for TF-IDF 0.9030
3 Logistic Regression for TF-IDF 0.8892
4 SVM for GloVe 0.80341246
5 Logistic Regression for GloVe 0.7700

Table 5.5: Models accuracy comparison table

Figure 5.6: Models Accuracy Comparison

28
Chapter 6

Concluding Remarks and Future

Work

6.1 Conclusion
Here, we’ll aim to investigate the possibilities for spotting bogus reviews that are
based on actual data by integrating active learning. In this work, various indica-
tors for spotting phony reviews are proposed and they are all given weights using
active learning. In the process, we train our model using active learning which iter-
atively learns from the best data. The review content’s TF-IDF values will be used
to build the feature vectors and classifiers like SVM and LR will be employed in
the classification process. Our experimental methods examine the precision of each
sentiment categorization algorithm and identify the technique that is most precise.
Additionally, by using detection techniques, we were able to identify phony negative
and good reviews. The bogus reviews are filtered using the unfair review detection
method. In this study, the SVM classifier outperformed the Logistic Regression in
terms of classification accuracy. Demonstrating how e↵ectively it predicts false re-
views examining the di↵erences in classifier accuracy was made easier by the data
visualization. Identifying false reviews is a difficult task. The absence of the tagged
dataset is the primary problem in this particular field of study. We suggested a
supervised method that does not require label information to predict the review
class in order to close this gap (fake or real). Research to identify unfair reviews
is extremely valuable in ensuring the reliability of reviews and giving customers a
positive purchasing experience. Makers can obtain accurate data by using this strat-
egy for detecting unjust reviews. By examining how customers feel about things,
businesses may monitor their product sales and reach. Customers have the option of
making a purchase or not. This technique so increases the credibility of e-commerce
websites. It has been noted that fake reviews are difficult to detect throughout this
study. Numerous studies have been conducted on this subject, but none of them
have produced a perfect conclusion (one hundred percent result). There are still
many flaws that are not being fixed, even in this day and age. We provided a way
for determining whether the provided comments on a specific good or service are
genuine or false using machine learning-based. Extending the dataset currently in
use and determining the best outcomes for a big quantity of data will o↵er insights
on performance in terms of accuracy as well as scalability. Also in future we will
try to give less value in dropout layer of BiDirectional LSTM and try to run more

29
epochs for better accuracy.

6.2 Limitations and Future Work

As we run our code in Google Colab free version so we could only run our code with
GPU for 5 hours constatntly after that it showed us the runtime session ended. So
we couldn’t run our LSTM model for long time which was needed. So in future we
will try to use Colab pro in future and get the better accuracy. Also for SVM and
LR using GloVe embedding we got less accuracy than TFIDF which shouln’t be the
case but we got this accuracy because we used average GloVe value for a sentence
words so we will try to use GloVe vectorizer with better approach which will give
us the better accuracy than TFIDF word embedding.

30
Bibliography

[1] A. Blum and T. Mitchell, “Combining labeled and unlabeled data with co-
training,” in Proceedings of the eleventh annual conference on Computational
learning theory, 1998, pp. 92–100.
[2] A. Z. Broder, “Identifying and filtering near-duplicate documents,” in Annual
symposium on combinatorial pattern matching, Springer, 2000, pp. 1–10.
[3] O. Chapelle, B. Schölkopf, A. Zien, et al., “Semi-supervised learning, vol.
2,” MIT Press, Cambridge. Cortes C, and Mohri M, et al.(2014) Domain
adaptation and sample bias correction theory and algorithm for regression.
Theoretical Computer Science, vol. 519, p. 103 126, 2006.
[4] B. Stein, M. Koppel, and E. Stamatatos, “Plagiarism analysis, authorship
identification, and near-duplicate detection pan’07,” in ACM SIGIR Forum,
ACM New York, NY, USA, vol. 41, 2007, pp. 68–71.
[5] N. Jindal and B. Liu, “Opinion spam and analysis,” in Proceedings of the 2008
international conference on web search and data mining, 2008, pp. 219–230.
[6] F. H. Li, M. Huang, Y. Yang, and X. Zhu, “Learning to identify review spam,”
in Twenty-second international joint conference on artificial intelligence, 2011.
[7] M. Ott, Y. Choi, C. Cardie, and J. T. Hancock, “Finding deceptive opinion
spam by any stretch of the imagination,” arXiv preprint arXiv:1107.4557,
2011.
[8] R. Y. Lau, S. Liao, R. C.-W. Kwok, K. Xu, Y. Xia, and Y. Li, “Text min-
ing and probabilistic language modeling for online review spam detection,”
ACM Transactions on Management Information Systems (TMIS), vol. 2, no. 4,
pp. 1–30, 2012.
[9] A. Morales, H. Sun, and X. Yan, “Synthetic review spamming and defense,” in
Proceedings of the 22nd International Conference on World Wide Web, 2013,
pp. 155–156.
[10] A. Mukherjee, V. Venkataraman, B. Liu, and N. Glance, “What yelp fake
review filter might be doing?” In Proceedings of the international AAAI con-
ference on web and social media, vol. 7, 2013.
[11] S. Shojaee, M. A. A. Murad, A. B. Azman, N. M. Sharef, and S. Nadali,
“Detecting deceptive reviews using lexical and syntactic features,” in 2013
13th International Conference on Intellient Systems Design and Applications,
IEEE, 2013, pp. 53–58.

31
[12] J. Li, M. Ott, C. Cardie, and E. Hovy, “Towards a general rule for identifying
deceptive opinion spam,” in Proceedings of the 52nd Annual Meeting of the
Association for Computational Linguistics (Volume 1: Long Papers), 2014,
pp. 1566–1576.
[13] M. I. Ahsan, T. Nahian, A. A. Kafi, M. I. Hossain, and F. M. Shah, “Review
spam detection using active learning,” in 2016 IEEE 7th Annual Information
Technology, Electronics and Mobile Communication Conference (IEMCON),
IEEE, 2016, pp. 1–7.
[14] D. Hovy, “The enemy in your own camp: How well can we detect statistically-
generated fake reviews–an adversarial study,” in Proceedings of the 54th An-
nual Meeting of the Association for Computational Linguistics (Volume 2:
Short Papers), 2016, pp. 351–356.
[15] E. I. Elmurngi and A. Gherbi, “Unfair reviews detection on amazon reviews
using sentiment analysis with supervised learning techniques.,” J. Comput.
Sci., vol. 14, no. 5, pp. 714–726, 2018.
[16] D. M. E.-D. M. Hussein, “A survey on sentiment analysis challenges,” Journal
of King Saud University-Engineering Sciences, vol. 30, no. 4, pp. 330–338,
2018.
[17] C.-C. Wang, M.-Y. Day, C.-C. Chen, and J.-W. Liou, “Detecting spamming re-
views using long short-term memory recurrent neural network framework,” in
Proceedings of the 2nd International Conference on E-commerce, E-Business
and E-Government, 2018, pp. 16–20.
[18] R. Bhargava, A. Baoni, and Y. Sharma, “Composite sequential modeling
for identifying fake reviews,” Journal of Intelligent Systems, vol. 28, no. 3,
pp. 409–422, 2019.
[19] C. G. Harris, “Comparing human computation, machine, and hybrid methods
for detecting hotel review spam,” in Conference on e-Business, e-Services and
e-Society, Springer, 2019, pp. 75–86.
[20] M. R. P. Kashti and P. S. Prasad, “Enhancing nlp techniques for fake review
detection,” Int. Res. J. Eng. Technol.(IRJET), vol. 6, pp. 241–245, 2019.
[21] G. Shahariar, S. Biswas, F. Omar, F. M. Shah, and S. B. Hassan, “Spam
review detection using deep learning,” in 2019 IEEE 10th Annual Information
Technology, Electronics and Mobile Communication Conference (IEMCON),
IEEE, 2019, pp. 0027–0033.
[22] Y. Fang, H. Wang, L. Zhao, F. Yu, and C. Wang, “Dynamic knowledge graph
based fake-review detection,” Applied Intelligence, vol. 50, no. 12, pp. 4281–
4295, 2020.
[23] E. Kau↵mann, J. Peral, D. Gil, A. Ferrández, R. Sellers, and H. Mora, “A
framework for big data analytics in commercial social networks: A case study
on sentiment analysis and fake review detection for marketing decision-making,”
Industrial Marketing Management, vol. 90, pp. 523–537, 2020.
[24] S. Saumya and J. P. Singh, “Spam review detection using lstm autoencoder:
An unsupervised approach,” Electronic Commerce Research, pp. 1–21, 2020.

32
[25] H. Tang and H. Cao, “A review of research on detection of fake commodity
reviews,” in Journal of Physics: Conference Series, IOP Publishing, vol. 1651,
2020, p. 012 055.
[26] S. N. Alsubari, S. N. Deshmukh, M. H. Al-Adhaileh, F. W. Alsaade, and T. H.
Aldhyani, “Development of integrated neural network model for identification
of fake reviews in e-commerce using multidomain datasets,” Applied Bionics
and Biomechanics, vol. 2021, 2021.
[27] P. Devika, A. Veena, E. Srilakshmi, A. R. Reddy, and E. Praveen, “Detec-
tion of fake reviews using nlp &sentiment analysis,” in 2021 6th International
Conference on Communication and Electronics Systems (ICCES), IEEE, 2021,
pp. 1534–1537.
[28] A. M. Elmogy, U. Tariq, M. Ammar, and A. Ibrahim, “Fake reviews detec-
tion using supervised machine learning,” International Journal of Advanced
Computer Science and Applications, vol. 12, no. 1, 2021.
[29] P. Gupta, S. Gandhi, and B. R. Chakravarthi, “Leveraging transfer learning
techniques-bert, roberta, albert and distilbert for fake review detection,” in
Forum for Information Retrieval Evaluation, 2021, pp. 75–82.
[30] N. Jnoub, A. Brankovic, and W. Klas, “Fact-checking reasoning system for
fake review detection using answer set programming,” Algorithms, vol. 14,
no. 7, p. 190, 2021.
[31] R. K. Kaliyar, A. Goswami, and P. Narang, “Fakebert: Fake news detection
in social media with a bert-based deep learning approach,” Multimedia tools
and applications, vol. 80, no. 8, pp. 11 765–11 788, 2021.
[32] R. Mohawesh, S. Tran, R. Ollington, and S. Xu, “Analysis of concept drift in
fake reviews detection,” Expert Systems with Applications, vol. 169, p. 114 318,
2021.
[33] R. Mohawesh, S. Xu, S. N. Tran, et al., “Fake reviews detection: A survey,”
IEEE Access, vol. 9, pp. 65 771–65 802, 2021.
[34] H. Paul and A. Nikolaev, “Fake review detection on online e-commerce plat-
forms: A systematic literature review,” Data Mining and Knowledge Discovery,
vol. 35, no. 5, pp. 1830–1881, 2021.
[35] S. R. Sahoo and B. B. Gupta, “Multiple features based approach for auto-
matic fake news detection on social networks using deep learning,” Applied
Soft Computing, vol. 100, p. 106 983, 2021.
[36] M. B. Alvi, M. Alvi, R. A. Shah, M. Munir, and A. Akhtar, “Machine learning-
based fake reviews detection with amalgamated features extraction method,”
Sukkur IBA Journal of Emerging Technologies, vol. 5, no. 2, pp. 10–17, 2022.
[37] H. Tufail, M. U. Ashraf, K. Alsubhi, and H. M. Aljahdali, “The e↵ect of fake
reviews on e-commerce during and after covid-19 pandemic: Skl-based fake
reviews detection,” IEEE Access, vol. 10, pp. 25 555–25 564, 2022.

[5] [19] [9] [7] [7] [12] [10] [8] [4] [2] [11] [10] [3] [6] [1] [20] [27] [23] [28] [14] [16] [30]
[37] [15] [24] [21] [13] [35] [31] [36] [29] [33] [28] [34] [22] [26] [17] [18] [32] [25]

33